Symptom

  • Even though specified as a single person or a single dog, multiple figures appear.
  • Parts like hands and faces seem to be multiplying.

Timing of occurrence

  • When generating with too high resolution such as 1024px or more in Stable Diffusion 1.5.
  • When generating with extreme vertical or horizontal resolution.

Cause

  • SD1.5 is trained on square images around 512px, so composition is difficult to stabilize at resolutions larger than that.

Solution

  • Generate at a size close to the model's recommended resolution
    • For SD1.5, try around 512-768px.